how to evaluate use model-00001-of-00002.safetensors and model-00002-of-00002.safetensors
Now I have finetune large-v3 and the resulting model is model-00001-of-00002.safetensors and model-00002-of-00002.safetensors. How can I use these two models for evaluation?
Hey
@andyweiqiu
- these are the sharded weights split by the .save_pretrained method. To use them for inference, simply follow the code snippets in the README, replacing model_id with the repo id or path to your save checkpoint: https://huggingface.co/openai/whisper-large-v3#usage
E.g. if I trained a model and saved it under sanchit-gandhi/whisper-large-v3-hi, I would set:
model_id = "sanchit-gandhi/whisper-large-v3-hi"
And keep the rest of the code example un-changed.
This is called speaker diarizatiom.
There are many repos on github for thid
Hey @andyweiqiu - these are the sharded weights split by the
.save_pretrainedmethod. To use them for inference, simply follow the code snippets in the README, replacingmodel_idwith the repo id or path to your save checkpoint: https://huggingface.co/openai/whisper-large-v3#usageE.g. if I trained a model and saved it under
sanchit-gandhi/whisper-large-v3-hi, I would set:model_id = "sanchit-gandhi/whisper-large-v3-hi"And keep the rest of the code example un-changed.
Thank you for your reply. I also want to know how to generate the pytorch_model.bin model file after I upload the local file(eg, model-00001-of-00002.safetensors and model-00002-of-00002.safetensors.) , or will it be generated automatically after uploading?
Hey
@andyweiqiu
- may I ask why you need the pytorch_model.bin format? Note that this weight format is inherently unsafe, as explained: https://huggingface.co/blog/safetensors-security-audit#why-create-something-new
Therefore, it is recommended to use weight sharding and safetensors serialisation to save PyTorch model weights: https://huggingface.co/docs/transformers/v4.35.2/en/main_classes/model#transformers.PreTrainedModel.save_pretrained.max_shard_size
These weights are entirely compatible with from_pretrained, so there's no need to change any of your code to accommodate for them!
You can use the same code-snippet as in: https://huggingface.co/openai/whisper-large-v3#usage
Just replace the model_id with the path (or repo id) of your model.
Hey @andyweiqiu - may I ask why you need the
pytorch_model.binformat? Note that this weight format is inherently unsafe, as explained: https://huggingface.co/blog/safetensors-security-audit#why-create-something-newTherefore, it is recommended to use weight sharding and safetensors serialisation to save PyTorch model weights: https://huggingface.co/docs/transformers/v4.35.2/en/main_classes/model#transformers.PreTrainedModel.save_pretrained.max_shard_size
These weights are entirely compatible with
from_pretrained, so there's no need to change any of your code to accommodate for them!You can use the same code-snippet as in: https://huggingface.co/openai/whisper-large-v3#usage
Just replace the
model_idwith the path (or repo id) of your model.
Ok, thank you. I've got it